src/amd/common/ac_gpu_info.c

   1 /*
   2  * Copyright © 2017 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining
   5  * a copy of this software and associated documentation files (the
   6  * "Software"), to deal in the Software without restriction, including
   7  * without limitation the rights to use, copy, modify, merge, publish,
   8  * distribute, sub license, and/or sell copies of the Software, and to
   9  * permit persons to whom the Software is furnished to do so, subject to
  10  * the following conditions:
  11  *
  12  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  13  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
  14  * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  15  * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
  16  * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  17  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  18  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  */
  25
  26 #include "ac_gpu_info.h"
  27 #include "sid.h"
  28 #include "gfx9d.h"
  29
  30 #include "util/u_math.h"
  31
  32 #include <stdio.h>
  33
  34 #include <xf86drm.h>
  35 #include <amdgpu_drm.h>
  36
  37 #include <amdgpu.h>
  38
  39 #define CIK_TILE_MODE_COLOR_2D                  14
  40
  41 #define CIK__GB_TILE_MODE__PIPE_CONFIG(x)        (((x) >> 6) & 0x1f)
  42 #define     CIK__PIPE_CONFIG__ADDR_SURF_P2               0
  43 #define     CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16          4
  44 #define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16         5
  45 #define     CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32         6
  46 #define     CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32         7
  47 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16    8
  48 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16    9
  49 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16    10
  50 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16   11
  51 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16   12
  52 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32   13
  53 #define     CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32   14
  54 #define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16   16
  55 #define     CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16  17
  56
  57 static unsigned cik_get_num_tile_pipes(struct amdgpu_gpu_info *info)
  58 {
  59    unsigned mode2d = info->gb_tile_mode[CIK_TILE_MODE_COLOR_2D];
  60
  61    switch (CIK__GB_TILE_MODE__PIPE_CONFIG(mode2d)) {
  62    case CIK__PIPE_CONFIG__ADDR_SURF_P2:
  63        return 2;
  64    case CIK__PIPE_CONFIG__ADDR_SURF_P4_8x16:
  65    case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x16:
  66    case CIK__PIPE_CONFIG__ADDR_SURF_P4_16x32:
  67    case CIK__PIPE_CONFIG__ADDR_SURF_P4_32x32:
  68        return 4;
  69    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x16_8x16:
  70    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_8x16:
  71    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_8x16:
  72    case CIK__PIPE_CONFIG__ADDR_SURF_P8_16x32_16x16:
  73    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x16:
  74    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x32_16x32:
  75    case CIK__PIPE_CONFIG__ADDR_SURF_P8_32x64_32x32:
  76        return 8;
  77    case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_8X16:
  78    case CIK__PIPE_CONFIG__ADDR_SURF_P16_32X32_16X16:
  79        return 16;
  80    default:
  81        fprintf(stderr, "Invalid CIK pipe configuration, assuming P2\n");
  82        assert(!"this should never occur");
  83        return 2;
  84    }
  85 }
  86
  87 static bool has_syncobj(int fd)
  88 {
  89         uint64_t value;
  90         if (drmGetCap(fd, DRM_CAP_SYNCOBJ, &value))
  91                 return false;
  92         return value ? true : false;
  93 }
  94
  95 bool ac_query_gpu_info(int fd, amdgpu_device_handle dev,
  96                        struct radeon_info *info,
  97                        struct amdgpu_gpu_info *amdinfo)
  98 {
  99         struct amdgpu_buffer_size_alignments alignment_info = {};
 100         struct amdgpu_heap_info vram, vram_vis, gtt;
 101         struct drm_amdgpu_info_hw_ip dma = {}, compute = {}, uvd = {}, vce = {}, vcn_dec = {};
 102         uint32_t vce_version = 0, vce_feature = 0, uvd_version = 0, uvd_feature = 0;
 103         uint32_t unused_feature;
 104         int r, i, j;
 105         drmDevicePtr devinfo;
 106
 107         /* Get PCI info. */
 108         r = drmGetDevice2(fd, 0, &devinfo);
 109         if (r) {
 110                 fprintf(stderr, "amdgpu: drmGetDevice2 failed.\n");
 111                 return false;
 112         }
 113         info->pci_domain = devinfo->businfo.pci->domain;
 114         info->pci_bus = devinfo->businfo.pci->bus;
 115         info->pci_dev = devinfo->businfo.pci->dev;
 116         info->pci_func = devinfo->businfo.pci->func;
 117         drmFreeDevice(&devinfo);
 118
 119         /* Query hardware and driver information. */
 120         r = amdgpu_query_gpu_info(dev, amdinfo);
 121         if (r) {
 122                 fprintf(stderr, "amdgpu: amdgpu_query_gpu_info failed.\n");
 123                 return false;
 124         }
 125
 126         r = amdgpu_query_buffer_size_alignment(dev, &alignment_info);
 127         if (r) {
 128                 fprintf(stderr, "amdgpu: amdgpu_query_buffer_size_alignment failed.\n");
 129                 return false;
 130         }
 131
 132         r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM, 0, &vram);
 133         if (r) {
 134                 fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram) failed.\n");
 135                 return false;
 136         }
 137
 138         r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_VRAM,
 139                                 AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 140                                 &vram_vis);
 141         if (r) {
 142                 fprintf(stderr, "amdgpu: amdgpu_query_heap_info(vram_vis) failed.\n");
 143                 return false;
 144         }
 145
 146         r = amdgpu_query_heap_info(dev, AMDGPU_GEM_DOMAIN_GTT, 0, &gtt);
 147         if (r) {
 148                 fprintf(stderr, "amdgpu: amdgpu_query_heap_info(gtt) failed.\n");
 149                 return false;
 150         }
 151
 152         r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_DMA, 0, &dma);
 153         if (r) {
 154                 fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(dma) failed.\n");
 155                 return false;
 156         }
 157
 158         r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_COMPUTE, 0, &compute);
 159         if (r) {
 160                 fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(compute) failed.\n");
 161                 return false;
 162         }
 163
 164         r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_UVD, 0, &uvd);
 165         if (r) {
 166                 fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(uvd) failed.\n");
 167                 return false;
 168         }
 169
 170         if (info->drm_major == 3 && info->drm_minor >= 17) {
 171                 r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_VCN_DEC, 0, &vcn_dec);
 172                 if (r) {
 173                         fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vcn_dec) failed.\n");
 174                         return false;
 175                 }
 176         }
 177
 178         r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_ME, 0, 0,
 179                                         &info->me_fw_version, &unused_feature);
 180         if (r) {
 181                 fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(me) failed.\n");
 182                 return false;
 183         }
 184
 185         r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_PFP, 0, 0,
 186                                         &info->pfp_fw_version, &unused_feature);
 187         if (r) {
 188                 fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(pfp) failed.\n");
 189                 return false;
 190         }
 191
 192         r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_GFX_CE, 0, 0,
 193                                         &info->ce_fw_version, &unused_feature);
 194         if (r) {
 195                 fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(ce) failed.\n");
 196                 return false;
 197         }
 198
 199         r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_UVD, 0, 0,
 200                                         &uvd_version, &uvd_feature);
 201         if (r) {
 202                 fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(uvd) failed.\n");
 203                 return false;
 204         }
 205
 206         r = amdgpu_query_hw_ip_info(dev, AMDGPU_HW_IP_VCE, 0, &vce);
 207         if (r) {
 208                 fprintf(stderr, "amdgpu: amdgpu_query_hw_ip_info(vce) failed.\n");
 209                 return false;
 210         }
 211
 212         r = amdgpu_query_firmware_version(dev, AMDGPU_INFO_FW_VCE, 0, 0,
 213                                         &vce_version, &vce_feature);
 214         if (r) {
 215                 fprintf(stderr, "amdgpu: amdgpu_query_firmware_version(vce) failed.\n");
 216                 return false;
 217         }
 218
 219         /* Set chip identification. */
 220         info->pci_id = amdinfo->asic_id; /* TODO: is this correct? */
 221         info->vce_harvest_config = amdinfo->vce_harvest_config;
 222
 223         switch (info->pci_id) {
 224 #define CHIPSET(pci_id, name, cfamily) case pci_id: info->family = CHIP_##cfamily; break;
 225 #include "pci_ids/radeonsi_pci_ids.h"
 226 #undef CHIPSET
 227
 228         default:
 229                 fprintf(stderr, "amdgpu: Invalid PCI ID.\n");
 230                 return false;
 231         }
 232
 233         if (info->family >= CHIP_VEGA10)
 234                 info->chip_class = GFX9;
 235         else if (info->family >= CHIP_TONGA)
 236                 info->chip_class = VI;
 237         else if (info->family >= CHIP_BONAIRE)
 238                 info->chip_class = CIK;
 239         else if (info->family >= CHIP_TAHITI)
 240                 info->chip_class = SI;
 241         else {
 242                 fprintf(stderr, "amdgpu: Unknown family.\n");
 243                 return false;
 244         }
 245
 246         /* Set which chips have dedicated VRAM. */
 247         info->has_dedicated_vram =
 248                 !(amdinfo->ids_flags & AMDGPU_IDS_FLAGS_FUSION);
 249
 250         /* Set hardware information. */
 251         info->gart_size = gtt.heap_size;
 252         info->vram_size = vram.heap_size;
 253         info->vram_vis_size = vram_vis.heap_size;
 254         /* The kernel can split large buffers in VRAM but not in GTT, so large
 255          * allocations can fail or cause buffer movement failures in the kernel.
 256          */
 257         info->max_alloc_size = MIN2(info->vram_size * 0.9, info->gart_size * 0.7);
 258         /* convert the shader clock from KHz to MHz */
 259         info->max_shader_clock = amdinfo->max_engine_clk / 1000;
 260         info->max_se = amdinfo->num_shader_engines;
 261         info->max_sh_per_se = amdinfo->num_shader_arrays_per_engine;
 262         info->has_hw_decode =
 263                 (uvd.available_rings != 0) || (vcn_dec.available_rings != 0);
 264         info->uvd_fw_version =
 265                 uvd.available_rings ? uvd_version : 0;
 266         info->vce_fw_version =
 267                 vce.available_rings ? vce_version : 0;
 268         info->has_userptr = true;
 269         info->has_syncobj = has_syncobj(fd);
 270         info->num_render_backends = amdinfo->rb_pipes;
 271         info->clock_crystal_freq = amdinfo->gpu_counter_freq;
 272         if (!info->clock_crystal_freq) {
 273                 fprintf(stderr, "amdgpu: clock crystal frequency is 0, timestamps will be wrong\n");
 274                 info->clock_crystal_freq = 1;
 275         }
 276         info->tcc_cache_line_size = 64; /* TC L2 line size on GCN */
 277         if (info->chip_class == GFX9) {
 278                 info->num_tile_pipes = 1 << G_0098F8_NUM_PIPES(amdinfo->gb_addr_cfg);
 279                 info->pipe_interleave_bytes =
 280                         256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX9(amdinfo->gb_addr_cfg);
 281         } else {
 282                 info->num_tile_pipes = cik_get_num_tile_pipes(amdinfo);
 283                 info->pipe_interleave_bytes =
 284                         256 << G_0098F8_PIPE_INTERLEAVE_SIZE_GFX6(amdinfo->gb_addr_cfg);
 285         }
 286         info->has_virtual_memory = true;
 287
 288         assert(util_is_power_of_two(dma.available_rings + 1));
 289         assert(util_is_power_of_two(compute.available_rings + 1));
 290
 291         info->num_sdma_rings = util_bitcount(dma.available_rings);
 292         info->num_compute_rings = util_bitcount(compute.available_rings);
 293
 294         /* Get the number of good compute units. */
 295         info->num_good_compute_units = 0;
 296         for (i = 0; i < info->max_se; i++)
 297                 for (j = 0; j < info->max_sh_per_se; j++)
 298                         info->num_good_compute_units +=
 299                                 util_bitcount(amdinfo->cu_bitmap[i][j]);
 300
 301         memcpy(info->si_tile_mode_array, amdinfo->gb_tile_mode,
 302                 sizeof(amdinfo->gb_tile_mode));
 303         info->enabled_rb_mask = amdinfo->enabled_rb_pipes_mask;
 304         memcpy(info->pa_sc_raster_config, amdinfo->pa_sc_raster_cfg,
 305                sizeof(info->pa_sc_raster_config));
 306         info->pa_sc_raster_config_1 = amdinfo->pa_sc_raster_cfg1[0];
 307
 308         memcpy(info->cik_macrotile_mode_array, amdinfo->gb_macro_tile_mode,
 309                 sizeof(amdinfo->gb_macro_tile_mode));
 310
 311         info->pte_fragment_size = alignment_info.size_local;
 312         info->gart_page_size = alignment_info.size_remote;
 313
 314         if (info->chip_class == SI)
 315                 info->gfx_ib_pad_with_type2 = TRUE;
 316
 317         return true;
 318 }
 319
 320 void ac_compute_driver_uuid(char *uuid, size_t size)
 321 {
 322         char amd_uuid[] = "AMD-MESA-DRV";
 323
 324         assert(size >= sizeof(amd_uuid));
 325
 326         memset(uuid, 0, size);
 327         strncpy(uuid, amd_uuid, size);
 328 }
 329
 330 void ac_compute_device_uuid(struct radeon_info *info, char *uuid, size_t size)
 331 {
 332         uint32_t *uint_uuid = (uint32_t*)uuid;
 333
 334         assert(size >= sizeof(uint32_t)*4);
 335
 336         /**
 337          * Use the device info directly instead of using a sha1. GL/VK UUIDs
 338          * are 16 byte vs 20 byte for sha1, and the truncation that would be
 339          * required would get rid of part of the little entropy we have.
 340          * */
 341         memset(uuid, 0, size);
 342         uint_uuid[0] = info->pci_domain;
 343         uint_uuid[1] = info->pci_bus;
 344         uint_uuid[2] = info->pci_dev;
 345         uint_uuid[3] = info->pci_func;
 346 }